1. No category

Hurdles and Steps:
Estimating Demand for Solar Photovoltaics
Kenneth Gillingham∗
Yale University and NBER
Tsvetan Tsvetanov†
Yale University
April 29, 2015
Abstract
This paper estimates demand for residential solar photovoltaic (PV) systems using a
new approach to address three empirical challenges that often arise with count data:
excess zeros, unobserved heterogeneity, and endogeneity of price. Our results imply
a price elasticity of demand for solar PV systems of -1.76. Counterfactual policy
simulations indicate that new installations in Connecticut in 2014 would have been 47
percent less than observed in the absence of state financial incentives, with a cost of
$135/tCO2 assuming solar displaces natural gas. Our Poisson hurdle model approach
holds promise for modeling the demand for many new technologies.
Keywords: count data; hurdle model; fixed effects; instrumental variables; Poisson;
solar photovoltaic systems; energy policy.
JEL classification codes: Q42, Q48, C33, C36
∗
Corresponding author: School of Forestry & Environmental Studies, Department of Economics, School
of Management, Yale University, 195 Prospect Street, New Haven, CT 06511, phone: 203-436-5465, e-mail:
kenneth.gillingham@yale.edu, and the National Bureau of Economic Research.
†
School of Forestry & Environmental Studies, Yale University, 195 Prospect Street, New Haven, CT
06511, phone: 650-521-2356, e-mail: tsvetan.tsvetanov@yale.edu.
1
Introduction
The market for rooftop solar photovoltaic (PV) systems has been growing rapidly around the
world in the past decade. In the United States, there has been an increase in new installed
capacity from under 500 MW in 2008 to over 4,500 MW in 2013, along with a decrease in
average (pre-incentive) PV system prices from over $8/W in 2008 to just above $4/W in 2013
(in 2014 dollars) (Barbose et al. 2014). These major changes in the market have come over a
period of considerable government support at both the state and federal levels, ranging from
state-level rebates to a 30 percent federal tax credit. For example, in Connecticut (CT), the
2008-2014 average combined state and federal incentives equaled 50 percent of the system
cost for most PV system purchasers.1 Yet such substantial government support is being
slowly reduced in many markets throughout Europe and the United States.
This paper makes both a methodological and substantive contribution. We develop a new
approach to addressing key empirical challenges in estimating the demand with count data.
We model the demand for residential PV systems in CT using rich installation-level data over
the period 2008 to 2014 to shed new light on a small, but fast-growing market that is similar to
many others around the world. Our estimate of the price elasticity of demand for a PV system
is -1.76. This estimate provides guidance to both policymakers and firms. Policymakers are
often interested in how changes in PV system prices – whether due to policy or other factors
– influence the sales of PV systems, for such knowledge is essential for assessing the impacts
of solar PV policies. Given the evidence that PV system installation markets are often
imperfectly competitive (Bollinger and Gillingham 2014; Gillingham et al. 2014b), firms may
be able use the elasticity to inform pricing decisions and market forecasts. Furthermore, this
estimate enables counterfactual policy simulations. We examine a counterfactual without
state financial incentives, finding that absent these incentives, installations in 2014 would
have been 47 percent less than observed, based on an estimated pass-through rate of 84
percent. We estimate the cost effectiveness of this incentive policy to be $135 per avoided
ton of CO2 if solar displaces natural gas generation.
The empirical challenges that motivate our approach are threefold: possible unobserved
heterogeneity at a fine geographic level, excess zeros, and endogeneity of price (due to simultaneity). Existing studies have developed count data methods that can separately address
the presence of excess zeros (e.g., Pohlmeier and Ulrich 1995; Santos Silva and Windmeijer 2001; Winkelmann 2004) or endogeneity of covariates without unobserved heterogeneity
(e.g., Mullahy 1997; Windmeijer and Santos Silva 1997; Terza 1998). Windmeijer (2008)
1
This calculation is made based on an average system price in 2008-2014 in CT of $36,607, average state
rebate amount of $10,288, and tax credit (taken post-incentive) of $7,896 for consumers with at least this
much tax burden (all in 2014 dollars).
1
goes further in addressing endogeneity in the presence of unobserved heterogeneity, but does
not address excess zeros. This paper is the first to address all three challenges in a count
data setting with clear policy significance. We use panel data on the count of annual solar
PV systems installed in a Census block group. Such panel data allow us to address unobserved heterogeneity in environmental preferences or other block group-specific factors with
fixed effects. However, 74 percent of the observations have zero values for the number of
installations in a block group in a year. Thus, classic models for count data, such as the
Poisson, are very likely to misspecify the underlying data generating process.
To address this problem, we employ a hurdle model, which is recognized as an effective
tool for dealing with the presence of excess zeros in count data settings (see, e.g., Cameron
and Trivedi 2013). We estimate a hurdle model based on two data generating processes:
a standard logit for whether a block group has at least one adoption and a zero-truncated
Poisson that models the rate of adoptions conditional on a block group having an adoption.
In order to tackle unobserved heterogeneity and endogeneity of the price variable, we extend
this model to accommodate fixed effects and instrumental variables.
At the basis of our approach is the conditional maximum likelihood (CMLE) estimator
for fixed effects logit models developed by Chamberlain (1980). Majo and van Soest (2011)
show that this conditional maximum likelihood approach can also be applied to the zerotruncated Poisson framework and present an application of this in a two-period setting. We
generalize the model of Majo and van Soest (2011) to a setting with an arbitrary number of
periods and possible endogeneity. In contrast to Majo and van Soest (2011), our approach
uses a generalized method of moments (GMM) estimator that is based on the first-order
conditions of the truncated Poisson CMLE. By using a GMM approach (as in Windmeijer
(2008) and several other papers), we can draw upon the established procedures for addressing endogeneity using valid instruments. To the best of our knowledge, this instrumental
variables Poisson hurdle model with fixed effects is new to the literature, and we confirm the
consistency of our estimator with a Monte Carlo simulation. This approach is particularly
useful for solar PV markets, but we expect that it is more broadly applicable to a many other
settings with similar empirical challenges, such as the demand for any early-stage technology.
Identification in our setting is based on deviations from block group means in both the
number of installations and PV system prices, after controlling for a variety of potential
confounders and instrumenting for price. As instruments, we use a set of supply shifters:
local electrical and wiring contractor wage rates and state incentives for PV systems. After
controlling for income at a localized level, county-year-level electrical and wiring wages act
as a valid supply shifter. Solar PV incentives in CT are given directly to the installing firm
rather than the consumer, so the consumer sees the post-incentive price at the bottom of
2
any contract to install a PV system. Thus, the incentives also act as a valid supply shifter.
Our preferred estimate of the price elasticity of PV system demand of -1.76 is the first
estimate we are aware of for CT. Using very different empirical strategies and data from
California, Rogers and Sexton (2014) find a rebate elasticity of approximately -0.4, while
Hughes and Podolefsky (2015) find an estimate of approximately -1.2. The elasticity we
obtain is somewhat higher than these two estimates, which may be due in part to differences
between the California and CT markets and in part to differences in estimation methodology.
Using our results, we also perform several policy counterfactual simulations to examine the
impact of state financial incentives and permitting policies on PV system adoption. We
first perform a simple analysis of the pass-through of the incentives to consumers, following
Sallee (2011), and find the pass-through rate of 84 percent mentioned above. Our simulation
results suggest that if we remove all financial incentives for purchasing PV systems, the
number of new installations in CT in 2014 would have been 47 percent percent less than
observed. This would result in up to 7.3 MW less added PV capacity in 2014. As mentioned
above, simple calculations suggest a cost-effectiveness of the program of $135 per avoided ton
of CO2 , assuming that solar power displaces natural gas-fired generation. Our simulations
also indicate that policies to eliminate permitting costs, such as those proposed in many
jurisdictions, would have increased the number of new installations in 2014 by 2.6 percent,
adding up to 0.4 MW of installed capacity.
The remainder of this paper is structured as follows. Section 2 provides a background on
policies in Connecticut that are targeted at stimulating solar demand. Section 3 describes
the data used in our analysis. Sections 4 and 5 outline the estimation methodology. Section
6 presents our empirical results. Section 7 presents the pass-through analysis, counterfactual
policy simulations, and a discussion of cost-effectiveness and welfare. Finally, Section 8
concludes.
2
Background on Solar PV Policies in Connecticut
Despite receiving fewer hours of sun than more southerly regions,2 CT has a robust and
growing market for solar PV systems, due to high electricity prices, many owner-occupied
homes, and considerable state support for PV systems. A $5 per watt (W) rebate for
residential solar PV systems (up to 10 kW) was available in CT as early as 2004. The
structure of incentives changed on July 1, 2011, with the passage of Connecticut Public Act
11-802, directing the newly established CT Energy Finance and Investment Authority, which
has since been renamed the Connecticut Green Bank (CGB), to develop a residential solar
2
See solar insolation maps at http://www.nrel.gov/gis/solar.html.
3
investment program that would result in at least 30 MW of new residential PV installations
by the end of 2022 (Shaw et al. 2014).3
Starting on March 2, 2012, two types of financial incentives were offered under the “Residential Solar Investment Program” (RSIP). For households that purchase a solar PV system,
CT offers an upfront rebate, under an “expected performance based buy-down” (EPBB) program (replaced by the similar “homeowner performance based incentive” (HOPBI) rebate
program after July 11, 2014). The rebate per W decreases based on the size of the system,
with lower incentives per W for systems larger than 5 kW and even lower for systems larger
than 10 kW. Most systems in CT fall under this incentive program until near the end of our
time period, when a much higher percentage of systems are not purchased outright. For such
third party-owned systems (e.g., installed under a solar lease or power purchase agreement),
CT offers quarterly incentive payments based on production (in kWh) over six years, under
a “performance-based incentive” (PBI) program (Shaw et al. 2014). All incentives are being
reduced over time in a series of steps as the solar PV market grows. Unlike in other states,
where learning-by-doing was an explicit policy motivation (Bollinger and Gillingham 2014),
this declining step schedule is primarily motivated in Connecticut by budget constraints and
a broad desire for a self-sustaining market. Table 1 provides details of the programs and the
step schedule.
Besides direct financial incentives, CT has several additional programs to promote solar
PV systems. One program is “net metering,” which allows owners of solar PV systems to
return excess generated electricity to the grid, offsetting electricity that is used during times
of non-generation, so the final electricity bill includes a charge for only the net electricity
usage. PV system owners can also carry over credits for excess production for up to one year.
On March 31 of each year, the utility pays the PV system owners any net excess generation
remaining at the avoided cost of wholesale electricity (a lower rate than the retail rate).
Another program in CT is an effort starting in 2011 to streamline municipality and
utility permitting, inspection, and interconnection processes. Several municipalities in CT
have reduced their permitting costs and streamlined their systems in response to this effort.
Similar efforts had been very effective at reducing permit fees in a number of other states,
such as Arizona, California, and Colorado, motivated in part by the substantially easier
permitting process in European countries such as Germany (CEFIA 2013). In CT there is
a proposal to either fully waive town permit fees or cap them at $200 per PV system. As of
January 2015, this legislative proposal is still under consideration.
But perhaps the most important non-rebate program for solar PV system adoption in CT
3
This target has already been reached in 2014, and proposals for expanding the program are underway.
See http://www.ctcleanenergy.com.
4
is the CGB-sponsored “Solarize CT” grassroots marketing program in selected municipalities.4 Solarize CT involves local campaigns with a municipality-chosen installer, roughly 20
week time frame, and pre-negotiated group discount pricing for all PV system installations in
the town. According to Gillingham et al. (2014a), this program led to a substantial increase
in demand in these municipalities: on average 35 additional installations per municipality
over the length of the program. The program is also associated with a roughly $0.40 to $0.50
per W decrease in installation prices.
3
Data
Our primary dataset contains nearly all residential solar PV system installations in CT from
the period 2008-2014. Each installation that receives a rebate in the two major investorowned electric utility regions in CT, United Illuminating and Eversource Energy (formerly
Connecticut Light & Power) is entered into a database by the CGB, with information on
price, rebates granted, system size (in kW), technical characteristics, financing arrangements,
address, date of application processing, and date of installation.5 Until late 2014, there were
few third party-owned (TPO) systems in CT (i.e., solar leases or power-purchase agreements)
and the price data for these TPO systems are well-known to be less reliable than ownedsystem price data. Thus, we exclude these TPO systems from our primary analysis.6 Our
raw dataset contains 5,070 residential PV installations approved for the rebate by the CGB
between January 1, 2008 and December 31, 2014.
We geocode the installations and match each to the U.S. Census block group they reside
in. We then collect U.S. Census data at the block group level from the 2006-2010, 2007-2011,
2008-2012, and 2009-2013 waves of the American Community Survey (ACS). We include data
on total population, median household income, median age, and education level. We obtain a
measure of block group population density by dividing population by land area. To generate
panel data for each variable, we use the average value for each variable across all waves
available for that year. So for example, the value for population in 2010 would be the average
of the 2006-2010, 2007-2011, 2008-2012, and 2009-2013 values for this variable. We use the
2009-2013 values for 2014.7 In addition to Census data, we also draw town-level voting
4
Solarize CT is funded by foundations, ratepayers, and other grants.
The only other utilities in CT are small municipal utilities in Bozrah, Norwich, and Wallingford (Graziano
and Gillingham 2014).
6
As a robustness check, we also re-do our analysis with the full dataset of purchased and third-partyowned (TPO) systems, controlling for third party ownership in the regressions. The results, presented in
Appendix F, should be interpreted with great caution due to the known issues with over-reporting of TPO
system prices.
7
We also employed an alternative approach, using only the mid-year of each ACS wave and interpolating
5
5
registration data from the Office of CT’s Secretary of the State (http://www.ct.gov/sots)
for each of the seven years in our study period. Finally, we bring in annual county-level
electrical and wiring wage data from the U.S. Bureau of Labor Statistics (http://www.bls.
gov).
3.1
Preparation of the Panel Dataset
Due to the possibility of unobserved heterogeneity at the localized level, we convert our data
to a panel dataset, where an observation is at the Census block group-year level. This leads
to a balanced panel of 10,150 observations in 1,450 block groups. For each block group-year,
we take the average of the system price, system size, and incentive level granted. We also
create an indicator variable for a Solarize campaign occurring in the given block group-year.
Using this panel dataset for our empirical analysis necessitates one further step. Many
observations refer to block group-year combinations where there are no installations. In
fact, our hurdle model approach is motivated by these excess zeros in our dataset. However,
we still have to determine the relevant installation price, system size, and incentive for
observations with no recorded contracts. This is a common issue in the empirical literature
and we take a conservative stance by examining two different approaches.
In our primary approach, we fill in the missing price and other variables with the average
annual value for the same municipality, and if this is not possible, we use the average annual
value within the county of the installation. Given that our adoption data in these block
groups is censored at zero, this approach may underestimate the price. Thus, we also examine
the results using the highest price, rather than the average price, in the above approach.
Fortunately, as shown in Section 6.2, our results are robust to this choice.
Our dependent variable, the post-incentive PV system price per W, is based on the ratio
of the average block group-year PV system prices and the average system size.
3.2
Trends and Summary Statistics
Figure 1 displays the overall trend in installations and PV system prices (in 2014 dollars)
during our study period. Between 2008 and 2014, the real average (post-incentive) price
falls by more than 40 percent, with a brief spike in 2009. This spike in the post-incentive
price is commonly attributed to a roughly 50 percent drop in the incentive between 2008
and 2009. Consistent with larger trends in the global PV market (Barbose et al. 2013), real
installation costs have been steadily decreasing during the entire study period. Figure 1 also
for all missing years, but this made no practical difference to our result. More generally, as shown in Section
6.2, our results are also robust to the exclusion of all demographic variables.
6
shows a substantial increase in PV system installations after 2011. Some of this increase
after 2011 consists of installations under Solarize programs, which involve both lower prices
and additional solar marketing.
Table 2 presents summary statistics for our panel dataset of 10,150 block group-year
observations. The number of PV system installations is a count variable, with a mean of
0.43 and a variance of 1.19. Recall that a Poisson model requires equidispersion, where the
mean equals the variance, so these data exhibit clear overdispersion. Figure 2 reveals why
this might be the case: the distribution of installations in a block group in a given year is
very strongly skewed, with most of the mass (over 74 percent) centered at zero. This visually
illustrates the issue of excess zeros in our dataset.
Table 3 presents summary statistics for the truncated dataset with positive installations.
Comparing Tables 2 and 3 reveals some differences. Most notably, there is a much stronger
presence of Solarize campaigns in the truncated sample (31 percent versus 13 percent). The
truncated sample also has a lower average population density, consistent with Graziano
and Gillingham (2014), who show that most installations in CT occur in suburban or rural
areas. The truncated sample also tends to have slightly larger systems, likely because it
more heavily samples block groups that are slightly wealthier and have a greater unobserved
preference for PV systems.
4
Empirical Specification: Preliminaries
Let Yit denote the number of PV installations in block group i purchased in year t. We use
W it to denote a vector containing the installation price pit and demand shifters. Demand
for solar is therefore represented by the following general function:
Yit = Dit (W it , Θ),
(1)
where Θ is a vector of parameters.
Before we describe the Poisson hurdle model, which is our preferred empirical specification, we first briefly review two alternative specifications: linear and Poisson. This exposition
lays the foundation for our Poisson hurdle model in Section 5.
4.1
Linear
In this specification, Dit (·) is modeled as a linear function and (1) becomes:
0
Yit = W it Θ + it ,
7
(2)
where it denotes an idiosyncratic error term. We can expand (2) further as follows:
0
Yit = αi + wit θ + µ(t) + it ,
where αi captures block group-specific effects, µ(t) is a polynomial in time,8 and wit contains
all remaining observed regressors.
The advantage of using a linear model is that it can easily accommodate fixed effects
to control for region-specific unobserved heterogeneity. Furthermore, the linear model also
allows for the straightforward implementation of an instrumental variables procedure to
address the endogeneity of prices. However, it is well-recognized in the literature (e.g., King
1988; Wooldridge 2002) that a linear model is unsuited for a count data setting. A linear
model not only misspecifies the count data generating process, but it also often predicts
negative and non-integer outcome values.
4.2
Poisson
In this specification, we model the data generating process as a Poisson process, which is
perhaps a better model for our count data. More specifically, if Yit is modeled as a Poisson
random variable with parameter λit > 0, its probability mass function, denoted by P o, is
given by:
λyit e−λit
, where y = 0, 1, 2, ..., ∞.
P o(y|λit ) = P r[Yit = y|λit ] =
y!
Furthermore, it can easily be shown that
E[Yit |λit ] = Var[Yit |λit ] = λit ,
(3)
i.e., the conditional mean and variance of the Poisson random variable are equal. The
parameter λ is modeled as an exponential function of the covariates:
0
λit = eW it Θ .
(4)
Therefore, in this specification, (1) becomes:
0
0
Yit = eW it Θ + it ≡ eαi +wit θ+µ(t) + it .
(5)
To facilitate discussion of this model, we can then further rewrite (5) using X to denote
8
An alternative to a flexible time trend is the inclusion of year fixed effects. We present year fixed effect
results in our robustness checks.
8
the vector of observed regressors and time trend variables, and δ to denote the vector of all
parameters besides the fixed effects parameters:
0
Yit = eαi +X it δ + it .
(6)
The presence of fixed effects usually poses a challenge in the estimation of nonlinear panel
models, such as the one presented in equation (6), since these individual-specific parameters
cannot be eliminated from the model through any of the standard transformations used in
linear models. If heterogeneity is left completely unrestricted and the model is estimated
directly with the fixed parameters included, the estimates would suffer from the “incidental
parameters problem” first noted by Neyman and Scott (1948). The problem arises since the
estimator of the fixed effects parameter αi is based only on the number of observations for
each i ∈ {1, ..., N }, and that number remains fixed even as N → ∞. Hence, the estimator
α
ˆ i is inconsistent, while the estimator of δˆ is only consistent if it can be expressed separately
from α
ˆ i . Fortunately, as shown by Blundell et al. (2002), the latter is the case in the Poisson
panel model. In particular, when this model is estimated by maximizing the log-likelihood
function, the first-order conditions can be rearranged so that δˆ does not depend on the
estimates of the fixed effects coefficients. Therefore, the maximum likelihood estimator
(MLE) for δ in our model with block group-specific constants does not suffer from the
incidental parameters problem.
Yet even with time-invariant unobserved heterogeneity accounted for, δˆM LE may still be
inconsistent due to the endogeneity of the price variable in W it . One potential solution to
the presence of an endogenous regressor in a Poisson model is the control function (CF)
approach. Imbens and Wooldridge (2007) lay out a CF approach for addressing endogeneity
in a cross-sectional Poisson model, which can be easily generalized to a panel data setting.
0
First, re-express the vector of covariates as W = (p, Z 1 ) and the corresponding vector of
0
parameters as Θ = (γ, Γ) . Furthermore, let Z be a vector of exogenous variables, such that
Z 1 ⊂ Z. Combining (3) and (4) and dropping all subscripts for simplicity, we can flexibly
re-write the conditional mean as:
0
E[Y |Z, p, ζ] = exp(γp + Z 1 Γ + ζ),
(7)
where ζ is an unobserved component of the conditional expectation that is correlated with
p. Let the endogenous variable p be written as:
0
p = Z Π + r.
9
(8)
Applying the law of iterated expectations to (7) and (8),
0
E[Y |Z, p] = E[Y |Z, r] = E[E[Y |Z, ζ, r]|Z, r] = exp(γp + Z 1 Γ)E[exp(ζ)|Z, r].
(9)
Note that both ζ and r are independent of Z. The standard assumption here is joint
normality of ζ and r, and under this assumption it can be shown that E[exp(ζ)|r] = exp(ρr)
for some scalar ρ. Hence, (9) becomes
0
0
E[Y |Z, p] = exp(γp + Z 1 Γ + ρr) = exp(W Θ + ρr),
which suggests a straightforward two-stage estimation procedure. In the first stage, we
estimate a linear regression of price on all excluded and included instruments in order to
obtain the residual rˆ. In the second stage, we estimate a Poisson model with W and rˆ as
covariates using MLE, as described earlier. Of course, since this is a two-stage procedure,
bootstrapping is perhaps the easiest way to obtain valid standard errors in the second stage.
The Poisson model can thus be used to address unobserved heterogeneity with fixed
effects and endogeneity of price. Moreover, it is likely a much better model for count data
than the linear specification. However, by virtue of its underlying distribution, the Poisson
model predicts that only a small proportion of all counts in the data equal zero and is likely
misspecified when there are excess zeros in the outcome variable.
5
Empirical Specification: Hurdle Model
5.1
General Form
In situations with very large numbers of zeros, the data generating process is perhaps better
thought of as a two-stage process. A hurdle model, originally formulated in Mullahy (1986),
is designed to model such a two-stage process.9 The first stage determines whether the count
variable has a zero or positive realization. A positive realization means that the “hurdle” is
crossed, in which case the exact outcome value is modeled by a truncated count distribution.
The two stages are functionally independent (Cameron and Trivedi 2013). For a given nonnegative count variable Y , let the first-stage process be driven by a distribution function f1 ,
9
A zero-inflated Poisson model is a similar model, but not as readily extendable to a panel data context
with fixed effects and instrumental variable estimation.
10
while the second stage governed by f2 . Then, the complete distribution of Y is given by:
(
f1 (0)
(1 − f1 (0)) f2 (y)
P r(Y = y) =
if y = 0,
if y = 1, 2, ..., ∞.
It is common to model the first stage using a logistic distribution, i.e., f1 = 1−P r[Y > 0],
where P r[Y > 0] is estimated through a logit regression model, and the second stage as a
P o(y)
zero-truncated Poisson, i.e., f2 = (1−P
(e.g., Min and Agresti 2005; Gallop et al. 2013;
o(0))
Neelon et al. 2013). Depending on the setting, the explanatory variables in the first- and
second-stage regressions may differ.
Adapting this framework to our setting,
0
eW it Θ1
P r[Yit > 0|W it , Θ1 ] =
1+e
0
W it Θ1
, f2 (y|λit ) =
λyit
y! (eλit − 1)
0
, and λit = eW it Θ2 .
Note that while we expect the same determinants to play a role in both stages of the
demand model, there is no reason to presume a priori that the effects of these determinants
would be identical. Therefore, we allow Θ1 to be different than Θ2 .
Let ιit = I(Yit > 0), where I(·) denotes the indicator function. The log-likelihood function
of this model can be expressed as follows:
L(Θ1 , Θ2 ) =
(
N X
T
X
(1 − ιit ) log
i=1 t=1
N X
T n
X
"
0
exp(W it Θ1 )
1
+ ιit log
0
0
1 + exp(W it Θ1 )
1 + exp(W it Θ1 )
#)
io
h
0
0
ιit Yit W it Θ2 − log(Yit !) − log exp(exp(W it Θ2 )) − 1
+
i=1 t=1
≡ Ll (Θ1 ) + Lt (Θ2 ),
(10)
where Ll (Θ1 ) is the log-likelihood function for the logit model and Lt (Θ2 ) is the log-likelihood
for the truncated Poisson model. Therefore, maximizing L(Θ1 , Θ2 ) is equivalent to maximizing the two terms separately, which implies that estimating the hurdle model effectively
reduces to separately estimating a binary logit model and a zero-truncated Poisson model.
This model is convenient both for estimation of the parameters and for estimation of the
price elasticity of demand. To see this, first note that the mean of the outcome variable in
11
this model is given by:
E[Yit |W it , Θ] = P r[Yit > 0|W it , Θ1 ]Et [Yit |W it , Θ2 ]
0
λit
exp(W it Θ1 )
,
=
0
1 + exp(W it Θ1 ) 1 − exp(−λit )
(11)
0
where Θ = (Θ1 , Θ2 ) and Et [·] is the mean in the truncated Poisson model. Let W =
0
(p, Z 1 ) , as in our earlier discussion. Then, the price elasticity η at the mean values of p and
Y can be derived through the following expression:
∂ [E[Yit |W it , Θ]]
E[pit ]
∂pit
E[Yit |W it , Θ]
∂ [P r[Yit > 0|W it , Θ1 ]]
E[pit ]
=
∂pit
P r[Yit > 0|W it , Θ1 ]
∂ [Et [Yit |W it , Θ1 ]]
Et [pit ]
≡ ηl + ηt .
+
∂pit
Et [Yit |W it , Θ1 ]
η =
In other words, η is a sum of the price elasticity from the logit and truncated Poisson
components of the hurdle model.
An important characteristic of this model is that, unlike the Poisson specification, it can
accommodate both overdispersion and underdispersion in the full dataset. In fact, it can be
shown that
Var[Yit |W it , Θ] = E[Yit |W it , Θ] + E[Yit |W it , Θ] (λit − E[Yit |W it , Θ]) .
Therefore, depending on the relative magnitudes of λit and E[Yit |W it , Θ], the model could
result in E[Yit |W it , Θ] ≤ Var[Yit |W it , Θ] or vice versa. In particular, E[Yit |W it , Θ] in (11)
is a linear function of P r[Yit > 0], so an excessive number of zero outcomes implies that
P r[Yit > 0] (and, hence, E[Yit |W it , Θ]) is relatively small, which leads to overdispersion
relative to the Poisson model.
Thus, the hurdle model is a promising approach for our setting. But the challenge of handling unobserved heterogeneity and endogeneity remains. While studies have extended the
traditional Poisson model to accommodate individual-specific fixed effects and instrumental
variables (e.g., Windmeijer 2008), there is little in the current literature on such extensions
of the hurdle model.
12
5.2
Fixed Effects
As suggested by equation (10), estimating the hurdle model is equivalent to separately estimating a logit model and truncated Poisson model. This allows us to address the challenge
of accommodating fixed effects separately in each of the two components of the hurdle model.
Logit
With fixed effects, the logit model becomes
0
0
exp(αil + wit θ 1 + µl (t))
bi exp(X it δ 1 )
,
P r[Yit > 0|W it , Θ1 ] =
≡
0
0
l
l
1 + exp(αi + wit θ 1 + µ (t))
1 + bi exp(X it δ 1 )
(12)
where bi ≡ exp(αil ) is a block group-specific coefficient, µl (t) is a polynomial time trend,
while X and δ 1 , as before, are used to denote the vectors of time variables and observed
regressors and their respective coefficients. As discussed in Section 4.2, the presence of N
block group-specific coefficients can be problematic in a nonlinear setting, unless δˆ1 can be
derived independently of ˆbi . For the case of logit with fixed effects, early work by Chamberlain
(1980) established that model parameters can be estimated using a conditional maximum
likelihood estimator (CMLE). In particular, this estimator uses the sum of outcomes within
each group i as a minimal sufficient statistic for i’s group-specific parameter. Using `li to
denote the conditional log-likelihood for all observations in block group i, it can be shown
that:
T
T
X
X
l
l
Yit ).
(13)
Yit ) = `i (δ 1 |
`i (bi , δ 1 |
t=1
t=1
In other words, the conditional log-likelihood does not depend on the fixed effects paramN
P
eter bi . The remaining parameters are then estimated by maximizing
`li . The conditional
i=1
maximum likelihood estimator for this model, δˆ1,CM LE , is consistent under mild restrictions
on the rate at which the sequence of bi ’s grows to infinity (Andersen 1970; Chamberlain
1980).
Truncated Poisson
The truncated Poisson model is estimated using only observations for which Yit > 0.
Allowing for unobserved block group heterogeneity, the truncated Poisson parameter λit for
each one of these observations can be expressed as follows:
0
0
λit = exp(αit + wit θ 2 + µt (t)) ≡ ci exp(X it δ 2 ) ≡ ci βit ,
13
(14)
with a time trend represented by µt (t), ci ≡ exp(αit ), δ 2 , as before, denoting a vector of pa0
rameters for the time variables and all remaining observed regressors, and βit ≡ exp(X it δ 2 ).
Unlike the fixed effects Poisson model, estimating a fixed effects zero-truncated Poisson
through maximum likelihood does not allow for δ 2 to be estimated independently of the
fixed effects coefficients. Therefore, MLE is inconsistent in this specification.
As shown by Majo and van Soest (2011), conditional maximum likelihood can be employed in this setting, in a similar fashion to the fixed effects logit model, which eliminates
the fixed effects parameters from the estimation. Majo and van Soest (2011) demonstrate
this method for a two-period panel dataset. In what follows, we generalize their procedure
to any panel with an arbitrary number of longitudinal observations.
0
Let Y i = (Yi1 , ..., YiT ) ∈ Y denote the T -dimensional vector of outcomes in block group i.
0
Also, let X i = (X i1 , ..., X iT ) denote the matrix of regressors. We now derive an expression
T
P
for the joint distribution of Y i conditional on X i ,
Yit , and the parameters. It can be
t=1
shown that
g(y|x, n, c, δ 2 ) = P r(Y i = y|X i = x,
T
X
t=1
T
1 Y n! yt
λ ,
Yit = n, c, δ 2 ) =
h t=1 yt ! t
(15)
where h is an expression that depends on the total number of observations with positive Yit
for a given i. In order to represent the general form of h, we simplify notation by dropping
the i subscript. Then, for any arbitrary T ∈ (1, T¯], where T¯ > 5, the general expression for
h takes the following form:
14
T
X
h(T, n, λ1 , ..., λT ) =
!n
−
λt
T
T
X
X
t=1
−I(T > 3)
+I(T > 4)
λt − λl
t=1
l=1
+I(T > 2)
!n
T X
T
X
T
X
l=1 m=1
m6=l
t=1
!n
λt − λl − λm
T X
T X
T
X
T
X
l=1 m=1 p=1
m6=l p6=l
p6=m
t=1
!n
λt − λl − λm − λp
T X
T X
T X
T
T
X
X
l=1 m=1 p=1 q=1
m6=l p6=l q6=l
p6=m q6=m
q6=p
¯
− . . . + I(T = T¯)(−1)T +1
!n
λt − λl − λm − λp − λq
t=1
T¯
X
λnt .
t=1
Note that, following from (14), and again dropping the i subscript for simplicity, we
have λt = cβt . So, λt is a linear function of c. This implies that h(T, n, λ1 , ..., λT ) =
cn h(T, n, β1 , ..., βT ). So, from (15),
T
T
Y n! y
Y n!
1
1
λt t = n
cyt βtyt
h(T, n, λ1 , ..., λT ) t=1 yt !
c h(T, n, β1 , ..., βT ) t=1 yt !
T
Y n! y
1
β t,
=
h(T, n, β1 , ..., βT ) t=1 yt ! t
i.e., g(y|x, n, c, δ 2 ) = g(y|x, n, δ 2 ), and, hence, the conditional distribution does not depend
on the nuisance parameter c. The conditional log-likelihood then takes the following form:
L(δ 2 ) =
N
X
i=1
(
N
X
`ti (δ 2 |
i=1
T
X
log[(
t=1
T
X
Yit ) =
t=1
Yit )!] −
T
X
log[Yit !] +
t=1
T
X
)
Yit log(βit ) − log(hi ) ,
(16)
t=1
where `ti denotes the conditional log-likelihood for all observations from block group i. Note
that since we condition on the sum of outcomes in a block group, only groups with more
than one positive outcome during the study period contribute to the log-likelihood.
15
This procedure is rather general and can be applied to an arbitrarily large number of longitudinal panel observations and any number of model parameters. The resultant parameter
estimator of this fixed effects zero-truncated Poisson model is obtained as δˆ2,CM LE = arg max
δ2
L(δ 2 ). In Appendix A, we follow the general proof of Andersen (1970) for CMLE with a
sufficient statistic for incidental parameters to show that the estimator δˆ2,CM LE is consistent
under similar assumptions to the ones imposed on the fixed effects logit model.
5.3
Endogeneity
Both the logit and truncated Poisson conditional likelihood estimators, discussed in Section
5.2, are consistent, provided that the respective models are specified appropriately. However,
this would no longer be the case with endogeneity of price. We thus extend the hurdle model
further in order to accommodate the implementation of a suitable instrumental variable procedure. Once again, we address the problem separately in the logit and truncated Poisson
portions of the model.
Logit
In the discrete choice literature, a number of methods have been developed to tackle endogeneity in demand settings. These include the product-market control (and instrumental
variable) approach of Berry et al. (1995), a simulated maximum likelihood approach developed by Gupta and Park (2009), and Bayesian methods employed by Yang et al. (2003)
and Jiang et al. (2009). More recently, Petrin and Train (2010) propose a control function
approach that is arguably easier to estimate and more flexible than the above methods, as it
does not require invoking equilibrium or imposing strict distributional assumptions for the
identification of demand parameters.
0
As in Section 4.2, dropping subscripts for simplicity, let W = (p, Z 1 ) and Z 1 ⊂ Z,
0
where Z is a vector of exogenous variables. Furthermore, let Θ1 = (γ1 , Γ1 ) . Suppose that
the purchase of any positive number of PV systems in a given block group generates utility
u, given by:
0
u = γ1 p + Z 1 Γ1 + ζ1 ,
(17)
where ζ1 denotes an idiosyncratic term that is correlated with price p. As in Section 4.2, let
0
p = Z Π + r.
(18)
Since ζ1 is correlated with p but not with Z, there exists some function CF (r, ρ1 ), where ρ1
16
is a parameter, such that
ζ1 = CF (r, ρ1 ) + ζ˜1 ,
and ζ˜1 is uncorrelated with p. We can then re-write (17) as
0
0
u = γ1 p + Z 1 Γ1 + CF (r, ρ1 ) + ζ˜1 = W Θ1 + CF (r, ρ1 ) + ζ˜1 .
(19)
Therefore, if we estimate our model including the control function CF (r, ρ1 ) with valid
instruments, we can consistently estimate Θ1 .
Petrin and Train (2010) suggest several simplifying assumptions. First, as an alternative
to specifying a joint distribution for both error terms in (17) and (18), one could enter r
flexibly in the utility and then choose a distributional assumption for ζ˜1 . Second, the control
function can be approximated as a linear function. Accordingly, we specify CF (r, ρ1 ) as
ρ1 r in (19) and assume that ζ˜1 is distributed i.i.d. type I extreme value.10 Therefore, the
probability of Y > 0 is now given by:
0
bi exp(X it δ 1 + ρ1 r)
.
P r[Yit > 0|W it , Θ1 ] =
0
1 + bi exp(X it δ 1 + ρ1 r)
As in our earlier discussion from Section 4.2, this CF specification implies a two-stage
estimation procedure. In the first stage, we run a linear regression of price on all excluded
and included instruments. The residuals from this stage are then included as a covariate in
the second-stage estimation, which, following the discussion from Section 5.2, is carried out
using a conditional maximum likelihood in order to eliminate the nuisance parameters. Finally, we bootstrap the standard errors to ensure that our inference accounts for both stages.
Truncated Poisson
To our knowledge, we are the first to address endogeneity of one or more of the regressors
in a fixed effects truncated Poisson model. In what follows, we develop a procedure that
stems from the truncated Poisson CMLE setup, derived in Section 5.2.
Suppose X it ∈ X ⊂ RP . Then, starting from (16), we can express the P -dimensional
score vector of derivatives of the log-likelihood corresponding to block group i as:
S i (δ 2 ) = ∇δ2 `ti (δ 2 |
T
X
t=1
Yit ) =
T
X
[Yit − φit (X i , ni , δ 2 )] X it ,
t=1
10
Alternatively, CF (r, ρ1 ) can be specified as a quadratic function (e.g., Olley and Pakes 1996). We also
re-estimated our model using a quadratic control function and found the results to be quite robust.
17
∂hi
βit
where φit ≡ ∂βithi .
P
Assuming that the model has been correctly specified,
g(y|x, n, δ 2 ) = 1 for all x, n,
y∈Y
and δ 2 , and it can be shown that the score of the log-likelihood function, evaluated at the
0 (p)
true parameter vector δ 02 , has a zero conditional
mean. Let Si (δ 2 ) denote the p-th element
of the score vector, corresponding to ∂δ∂p `i (δ 2 )
. Then,
2
0
δ 2 =δ 2
X ∂
=
g(y|X i , ni , δ 02 ) =
p `i (δ 2 )
∂δ2
δ 2 =δ 02
y∈Y
!
X ∂
X
∂
=
g(y|X
,
n
,
δ
)
= 0.
g(y|X
,
n
,
δ
)
i
i
2
i
i
2
p
p
∂δ
∂δ
0
2
2
δ 2 =δ 2
δ 2 =δ 02
y∈Y
y∈Y
E[Si (δ 02 )(p) |X i , ni ]
Since the above is true for any element of the score, it follows that
E[S i (δ 02 )|X i , ni ] = 0.
(20)
Clearly, this result would not hold if one or more of the variables in X are endogenous in
the model. Let
ξit = Yit − φit (X i , ni , δ 02 ),
(21)
0
and let ξ i = (ξi1 , ..., ξiT ) . Thus, by the law of iterated expectations, (20) implies that
0
E[X i ξ i ] = 0,
(22)
provided that X i is exogenous.
Note that (21) and (22) can be viewed as a reduced-form representation of our demand
model, in which ξit is an error term that is orthogonal to the vector of regressors, as long as
these regressors are exogenous in the model. If exogeneity of X holds, the sample analog
of (22) is equivalent to the first-order conditions of the conditional likelihood from (16) and
yields δˆ2,CM LE which was shown to be a consistent estimator of δ 02 .
However, if X is not exogenous in the demand model, (22) no longer holds and the
estimator δˆ2,CM LE is inconsistent. In that case, a vector Z, comprised only of variables that
are exogenous in the model, would still be orthogonal to the error term in (21), i.e.,
0
E[Z i ξ i ] = 0,
0
(23)
where Z i = (Z i1 , ..., Z iT ) . We can then use the sample analog of (23) in order to estimate
δ 02 through a generalized method of moments (GMM) estimator, provided that there are a
18
total of at least P exogenous variables in Z. Suppose Z it ∈ Z ⊂ RQ , where Q > P , and let
0
ψ(Z i , δ 2 ) = Z i ξ i . Then,
"
δˆ2,GM M
N
1 X
ψ(Z i , δ 2 )
= arg max
δ 2 ∈∆ N
i=1
#0
#
N
X
1
ˆ
ψ(Z i , δ 2 ) ,
Ξ
N i=1
"
(24)
ˆ is an optimal weighting matrix. As shown in Appendix A, under the standard
where Ξ
GMM assumptions, δˆ2,GM M is a consistent estimator of δ 02 .
5.4
Deriving the Price Elasticity
The estimators derived in Sections 5.2 and 5.3 are all based on conditioning out the incidental parameters in the model. Since we do not estimate all of the parameters in the
original model specification, we are not able to calculate exact marginal effects, needed for
the derivation of price elasticity. Instead, we develop, in the spirit of Kitazawa (2012), a
procedure for obtaining average elasticity estimates that is shown to converge to the true
average elasticity values as N → ∞.
Logit
The average price elasticity in the logit model is given by:
ηl = (1 − P r[Yit > 0|W it , Θ1 ]) γ1 E[pit ],
(25)
where γ1 is the estimated coefficient on price.
0
b exp(X δ )
Note that P r[Yit > 0|W it , Θ1 ] = 1+bi exp(Xit0 1δ ) . Since we do not have an estimate of bi ,
i
it 1
we cannot use the post-estimation predicted probabilities to calculate ηl . However, by the
N P
T
P
Law of Large Numbers, ¯ι = N1T
ιit is a consistent estimator of P r[Yit > 0|W it , Θ1 ].
p
i=1 t=1
Similarly, p¯ →
− E[pit ], where p¯ is the sample average of price. Therefore, we can obtain a
consistent estimate of ηl through the following expression:
ηˆl = (1 − ¯ι)ˆ
γ1 p¯,
(26)
where γˆ1 can be obtained through a CF approach in our setting.
Truncated Poisson
In a similar vein, we derive an estimator for the average elasticity in the zero-truncated
19
Poisson portion of the hurdle model. First, it can be shown that ηt is given by
ηt = (1 − exp(−λit )Et [Yit |W it , Θ2 ]) γ2 Et [pit ].
(27)
By re-arranging the expression for Et [Yit |W it , Θ2 ], we obtain
− exp(−λit ) =
λit
− 1.
Et [Yit |W it , Θ2 ]
Hence, (27) becomes
ηt = (1 + λit − Et [Yit |W it , Θ2 ]) γ2 Et [pit ].
Similar to our earlier discussion, Y¯t and p¯t , the means of Y and p from the truncated Poisson
sample, are consistent estimators of Et [Yit |W it , Θ2 ] and Et [pit ], respectively.
0
We also need to derive a separate estimator of λ, since λit = ci exp(X it δ 2 ) implies that
we cannot calculate λ from the estimated parameters as we do not estimate ci . Let
Et [Yit |W it , Θ2 ] =
λit
≡ m(λit ).
1 − exp(−λit )
Then, λit = m−1 (Et [Yit |W it , Θ2 ]). As shown in Appendix B, m(· ) is a monotonic function
over the relevant range of λit values in this model, implying that m−1 (· ) is a one-to-one
p
mapping from Et [Yit |W it , Θ2 ] to λit . Furthermore, since Y¯t →
− Et [Yit |W it , Θ2 ], it follows
that m−1 (Y¯t ) is a consistent estimator of λit . Hence,
ηˆt = 1 + m−1 (Y¯t ) − Y¯t γˆ2 p¯t ,
where γˆ2 is obtained through GMM in our setting.
5.5
Monte Carlo Simulations
To evaluate the performance of the Poisson hurdle estimator, we conduct a set of Monte Carlo
simulations. In short, we simulate data from our data generating process and then apply
our estimator and alternative estimators to the simulated data. We use a simple panel data
framework, in which the outcome count variable is determined by a set of unobserved timeinvariant and time-varying cross-section-specific factors and a single endogenous regressor.
The model, data generation process, and parameter values, used in these simulations, are
described in Appendix C.
Our results are encouraging. Table 4 shows the results from one representative simulation.
The bias on each of the parameters of interest using the logit CF and truncated Poisson GMM
20
is small: 3 percent for the logit CF and 0.2 percent for the truncated Poisson. Converting
this to an elasticity, we find that our estimated (combined) elasticity is within 1.2 percent
of true elasticity value. Comparing these results from using linear least squares GMM and
Poisson CF approach, we find that our estimated elasticity is much closer to the true value.
We repeat this procedure with several different sets of parameter values and find similar
results.
6
6.1
Results
Primary Results
Our identification strategy relies on the validity of our instruments. As described earlier, we use two marginal cost shifters as instruments for the post-incentive price: average
EPBB/HOPBI incentive levels in $/W for the first 5 kW of installed capacity in each block
group-year and county-year average electrical/wiring wages. Incentives are given directly to
installers in CT, so they act as a shifter of the firm marginal cost. County-year average electricial/wiring wages are used as a proxy for PV system installer wages, and are a valid shifter
after controlling for income. Appendix D contains the results of the first-stage regression,
demonstrating that these are strong instruments.
Table 5 presents the results from estimating a linear model, Poisson model, and Poisson
hurdle model with instruments for price.11 Our preferred specification is the instrumental
variables hurdle model, consisting of a logit regression estimated with a control function
approach in column (3) and a trucated Poisson estimated by GMM in column (4).12 All
columns include block group fixed effects and a quadratic time trend. Standard errors are
clustered at the town level. At the bottom of the table, we present estimates of the price
elasticity of demand for each specification.
We are most interested in the statistical and economic significance of the price coefficient.
We find negative and statistically significant coefficients on the price variable in all model
specifications. Rather than interpret the coefficient directly, we find it more instructive to
consider the price elasticity implied by the coefficient taken at the mean of our sample.13
11
For completeness, Appendix E also presents output from the same model specification without the use
of instrumental variables. We view these results with caution as the coefficients are most likely biased due
to the endogeneity of price.
12
While over-identification tests are never definitive, we use a Hansen’s J over-identification test to examine the validity of the instruments in the truncated Poisson GMM specification of our hurdle model.
Comfortingly, we find that we fail to reject the null of valid instruments, with a Chi-squared test statistic of
0.56 and a p-value of 0.46. We find similar results for other specifications.
13
The price elasticity in the linear model is estimated as η = γˆY¯p¯ , where γˆ is the two-stage least squares
21
Estimating the model using a count specification implies a considerably higher (in absolute
value) price elasticity relative to the linear specification. A linear 2SLS regression yields a
price elasticity of -0.51. Fitting a Poisson model generates an elasticity coefficient that is
almost four times higher. However, both of these empirical specifications are unsuited for
our data with excess zeros, as discussed in Section 4.
As shown earlier, the total price elasticity of the hurdle model, which is our preferred
specification, is the sum of the elasticities from the logit and truncated Poisson. Thus, the
hurdle model implies a price elasticity of solar PV system demand of -1.76 at the mean of our
sample. At the average system price and number of installations in our sample, this elasticity
estimate suggests that a $1/W decrease in the installation price (well within the variation
in our data) would lead to an increase in demand of approximately 0.7598 additional PV
systems in each block group during the respective year. In our data, there are 1,450 block
groups in any given year. Thus, a $1/W decrease in system price translates into 1,102
additional installations demanded statewide in that year.
The other coefficients are of less interest to us, but we are reassured to see that the signs
are generally consistent and make sense. One of the more interesting of these is the coefficient on the indicator for whether a block group has a Solarize campaign, which is positive
and highly statistically significant in all specifications except the logit. This is consistent
with results in Gillingham et al. (2014a), which use quasi-experimental and experimental
approaches to find a large treatment effect of the Solarize program. Furthermore, given the
structure of Solarize programs, which rely on word of mouth and spread of information between adopters and non-adopters, it is intuitive that the presence of a campaign in a block
group is likely to have a positive impact on demand conditional on some adoptions having
taken place during the respective year. The other coefficients largely make sense in sign.
Solar demand is higher in less densely populated areas, in areas with a younger population,
and in areas with a higher percentage of college graduates (only statistically significant in the
truncated Poisson). The general lack of statistical significance for these demographic variables is likely attributable to the lack of substantial time series variation in these variables,
given that the block group fixed effects absorb the cross-sectional variation.
6.2
Robustness Checks
We perform several robustness checks, which are both reassuring and provide insight into
the variation behind our results. Table 6 shows the estimated elasticities from each of these
(2SLS) estimate of the price coefficient and p¯ and Y¯ are the sample means for price and number of installations. The price elasticity in the Poisson model is estimated as η = γˆ p¯, where γˆ is the Poisson CF estimate
of the price coefficient.
22
robustness checks and includes the elasticity estimates from our preferred specifications in
columns (3) and (4) of Table 5 for comparison.
First, we are concerned that our results in the logit regression may be driven by the
method we use to fill in missing price data in block group-installer-years where no contracts
were signed. We therefore re-estimate the model using the highest, rather than average,
recorded prices as proxies, preserving the order of our approach outlined in Section 3.1
(column I). We find that our logit CF estimates are hardly affected by this change in the
interpolation approach, a very reassuring result.
Next, we test the extent to which our results are affected by the use of year fixed effects
instead of a quadratic time trend. While potentially confounding changes over time should
be relatively smooth over this time period, we feel that it is useful to compare our primary
results to those using variation stemming from deviations from the mean in each year and
block group. However, given that we only have seven years in our sample, we not only
have small T , potentially leading to a large finite sample bias, but would also lose a great
deal of the useful identifying variation in our data. Our results in column II indicate that
including year fixed effects does not leave sufficient variation in the data to identify the price
coefficients.
Third, we test the sensitivity of our results to an alternative specification without demographic and voting variables among the regressors (column III). Excluding these variables,
we obtain almost identical elasticity estimates as in our baseline logit and truncated Poisson
runs. This result is not surprising, given the lack of substantial temporal variation in these
variables, noted earlier.
Lastly, we examine the effect of excluding observations from Solarize campaigns from
the analysis (column IV). This robustness check also provides insight into the origins of our
results. We find an elasticity estimate from the truncated Poisson GMM that is of smaller
magnitude than the baseline value, while the logit CF is much larger. The overall implied
elasticity from the hurdle model is quite close to our baseline elasticity.
These results underscore the importance of the Solarize campaigns for our observations
with multiple installations in a block group-year. Figure 3 shows a histogram of observations
with at least one installation. Non-Solarize observations with positive outcomes are largely
centered at 1, while higher-count outcomes are mostly Solarize observations. Hence, when
we drop all Solarize observations, our outcome variable is effectively reduced to a binary
variable, enabling us to capture most the variation in the data through the logit component
of the hurdle model. This suggests that the hurdle model, as a mix of components that can
exploit both binary and zero-truncated count data variation, offers a flexible approach for
estimating solar demand in various settings: from emerging markets with few installations
23
to booming higher-demand markets.
7
Policy Analysis
In this section, we highlight what our results mean for policies in the solar PV market
in CT. We run four policy counterfactuals: a removal of all state financial incentives, a
reduction in the state financial incentives down to Step 6 (recall Table 1), a reduction in
state incentives to half of their average level in 2014, and a waiving of all permitting fees for
solar PV systems. These policy simulations are particularly relevant, for CT plans to phase
out financial incentives entirely in the next few years and at the same time has made efforts
to convince municipalities to reduce or entirely waive permitting fees (CEFIA 2013). In
addition, we calculate the rate of free-ridership and the cost-effectiveness in terms of dollars
per ton of avoided carbon dioxide emissions for the financial incentive policy.
7.1
Simple Pass-through Analysis
As a first step, the above analysis necessitates obtaining a measure of the pass-through
of installation costs to customers. While there are some estimates of pass-through in the
California solar PV system market, the CT market may be quite different than the California
market. Moreover, these estimates based on California data vary widely. For example,
Henwood (2014) finds a low pass-through rate, while Dong et al. (2014) find nearly complete
pass-through. These working papers use similar data, but very different empirical strategies.
Thus, we perform our own simple pass-through analysis, following the standard approach
for estimating pass-through of a subsidy or tax on consumer prices (see, e.g., Sallee 2011).
We observe the rebate received by each of the 5,070 purchased installations in 2008-2014.
Due to the declining incentives and time lag between the submission and approval of contract
applications, there is both temporal and cross-sectional variation in rebate levels. We exploit
this variation in order to identify the effect of rebates on pre-rebate price, after controlling
for unobserved region-specific and time effects.
Using installation-level data, we estimate a linear regression of pre-rebate price (in $/W)
on the rebate level for the first 5 kW of installed capacity (in $/W), a Solarize campaign
indicator variable, a quadratic weekly trend, and municipality fixed effects. Our results are
shown in Table 7.14 The estimated coefficient on the rebate variable can be interpreted
as the fraction of a $1/W increase in the rebate level that is captured by the installer. A
14
We also re-ran the estimation, aggregating all data at the block group-year level, replacing municipality
fixed effects with block group fixed effects, and using yearly instead of weekly time trend. We obtained an
almost identical rebate coefficient.
24
coefficient equal to zero would suggest complete pass-through: the rebate is entirely captured
by consumers. A coefficient equal to one would suggest zero pass-through: the rebate is
entirely captured by firms. We find a highly statistically significant coefficient on the rebate
level of 0.16, implying a pass-through rate of 84 percent. In other words, a $1/W decrease in
installer costs translates into an 84 cents/W decrease in the PV system price. Such a passthrough rate is indicative of some imperfect competition in the solar PV market, consistent
with findings in Bollinger and Gillingham (2014) and Gillingham et al. (2014b).
7.2
Policy Simulations
For each policy scenario, we quantify the counterfactual number of installations in CT in
2014 and 2015, and compare these to the observed number of installations in 2014 and a
projected number of installations in 2015. We limit our simulations to this two-year period
in order to ensure that the demand structure and system characteristics are relatively stable.
In this rapidly changing market, extrapolating too far out is not likely to be a useful exercise.
Our methodology is straightforward. The CGB database contains installer-reported data
on the module, inverter, permitting, and labor costs for each installation.15 Since the statutory incidence of the policies falls on the installing firms, we model the policy counterfactuals
as a change in the installer’s marginal cost. We find the percentage change in the sum of the
module, inverter, permitting, and labor costs due to the policy, adjust this by a pass-through
rate to estimate the change in price, and then use our estimated elasticity (-1.76) to calculate
the percent change in the number of installations. This is then converted into total number
of new installations and additional installed capacity in MW. Further details of our approach
used for the counterfactual simulations are in Appendix G.
The results of our four policy counterfactual simulations are shown in Table 8. In the
baseline scenario, state policies are assumed to remain unchanged and system prices are
affected only by the falling global costs of PV components. Hence, projected installations
and added capacity for 2015 under this scenario are slightly higher relative to the observed
numbers in 2014. In the “Small Reduction” counterfactual, the rebate offered by the state is
reduced down to Step 6 in the EPBB/HOPBI schedule in Table 1, lowering the average incentive amount in 2014 from the observed $0.935/W to $0.675/W. In the “Moderate Reduction”
counterfactual, state incentives are cut in half, resulting in an average rebate of $0.468/W,
whereas in the “Phase-out” counterfactual, incentives are reduced to zero beginning January
2014.16 Finally, the “Permit Fees” counterfactual simulates a decrease in average PV system
15
We are not entirely confident in these reported cost estimates in the dataset, so we also use the module
and inverter price indices instead of the reported module and inverter costs. We find very similar results.
16
The variation in prices in our data is on the order of the change in prices from this counterfactual, giving
25
costs from waiving all municipal permit fees. This amounts to a reduction in average system
costs by approximately $0.052/W in 2014.
Our simulations suggest that a reduction of incentive levels to Step 6 would have resulted
in a 13.1 percent decrease in the number of new PV installations in CT during 2014 relative
to the number observed, and 13.2 percent in 2015. This would imply 258 fewer installations
in 2014 and 268 fewer in 2015, and is equivalent to a reduction in added PV capacity of 2.01
MW in 2014 and 2.09 MW in 2015. Under a moderate incentive reduction, the simulated
impacts on new installations and capacity are about 80 percent higher compared the “Small
Reduction” scenario. Furthermore, had there not been any state financial incentives, the
number of new installations would have been 47 percent less than observed in 2014 and 47.3
percent less in 2015. This is a difference of 927 installations in 2014 and 957 installations in
2015. In terms of impact on added PV capacity, the reduction would have been up to 7.25
MW in 2014 and 7.49 MW in 2015.
Finally, if CT succeeded in eliminating municipality permit fees, the number of installations would have been increased by 2.6 percent in 2014 and 2.7 percent in 2015. This
amounts to 51 additional installations in 2014 and 54 additional installations in 2015, with
an associated increase in added capacity of 0.40 MW in 2014 and 0.42 MW in 2015. Of
course, this may be an underestimate of the effect of eliminating permitting fees, since it
does not account for the reduced paperwork and associated labor costs from expedited permitting, which likely would occur with a reduction or elimination of the fees. Such additional
cost reductions may also be passed on to consumers, leading to a greater effect.
7.3
Cost-Effectiveness of Subsidies
Our above results can be used to provide insight into the cost-effectiveness and welfare
implications of the subsidy policy in CT. First, our “No subsidy” counterfactual policy immediately provides insight to the level of free-ridership from the policy. The results indicate
that roughly half of those who installed in 2014 and 2015 would have installed anyway,
indicative of a moderate rate of free riding.
What does this mean for the cost-effectiveness of the policy? A quick calculation based
on our results reveals that the public cost of the subsidy program is $1.12 per additional watt
installed (in 2014 dollars). To determine the benefits from a watt of installed solar capacity,
we first calculate the average annual electricity generation from a PV system in CT. We
use the National Renewable Energy Laboratory’s “PVWatts” solar electricity calculator for
four different locations across CT and average the resultant output.17 We assume a 25-year
us confidence that we are not extrapolating too far out-of-sample.
17
See http://pvwatts.nrel.gov. The locations we use are New London, North Canaan, Putnam, and
26
lifespan of an average installed system, which is in line with both the standard warranty
offered by most manufacturers18 and the assumptions made by the CGB in calculating
lifetime electricity generation for an average system. This yields an average lifetime electricity
generation of 32.257 kWh per watt of installed capacity, implying a public cost of the program
of $0.035/kWh generated.
We can also calculate the cost-effectiveness in terms of environmental benefits. There are
two sources of avoided pollution from installing solar PV systems: greenhouse gas emissions
and local air pollutants. The amount of avoided pollution depends on assumptions about
the type of generation that is displaced by solar PV both now and for the next 25 years.
The average carbon dioxide emissions rate in the Northeast is estimated to be 0.000258 tons
of CO2 per kWh (Graff Zivin et al. 2014). Assuming this holds for the next 25 years, the
lifetime cost-effectiveness of the program would be $135/tCO2 . If we assume that the CT
electricity grid will continue to become less carbon intensive over time, this estimate will be
correspondingly higher.
This back-of-the-envelope calculation does not include the benefits from reducing local
air pollution, which also impact social welfare. We can get some sense of the welfare impacts
of the subsidy policy by employing estimates of external damages from criteria air pollutants
from Muller et al. (2011). We use these and estimates of the social cost of carbon of $37/tCO2
(in 2007 dollars) from IAWG (2013) to estimate the total external costs from natural gas-fired
generation to be $0.021/kWh (in 2014 dollars). From coal-fired generation, this estimate is
$0.051/kWh (in 2014 dollars). Since coal provides only a small and decreasing fraction of the
electricity generation, the natural gas estimate is most useful. Comparing this $0.021/kWh
estimate of external costs to the public cost of the program of $0.035/kWh, suggests that
solar PV has been over-subsidized in CT. Of course, there are numerous caveats to this simple
calculation, including the social cost of public funds, intermittency issues (Gowrisankaran
et al. 2015), uncertainty in future electricity generation, and possible spillovers in learningby-doing at the localized level (Bollinger and Gillingham 2012). Yet, we do find it suggestive,
and consistent with estimates in California (Borenstein 2012; Rogers and Sexton 2014).
8
Conclusions
This study estimates the demand for solar PV systems using a new empirical approach: a
Poisson hurdle model with fixed effects and instrumental variables. This approach allows
Stamford. The calculations account for reduction in output due to periods of no sunlight, as well as loss of
generation during conversion from direct to alternating current power.
18
See, e.g., http://energyinformative.org/solar-panel-warranty-comparison.
27
us to tackle several key challenges that arise in modeling count data in the diffusion of
any new technology. Specifically, it addresses unobserved heterogeneity at a fine geographic
level, excess zeros in the outcome variable, and the endogeneity of price. In addition to the
adoption of new technologies, we also expect this approach to be useful in a variety of other
settings, such as the demand for health care in hospital units.
We estimate the price elasticity of demand for solar PV systems in CT over 2008-2014
to be -1.76. This estimate is valuable to both policymakers and firms. As module prices
continue to drop, it provides useful guidance for forecasting the number of new installations,
absent policy changes. It is also very useful for examining changes in policy, in light of
continuing policy discussions about phasing out the financial incentive program and reducing
municipal permit fees. After estimating a pass-through rate of 84 percent, we perform several
counterfactual policy simulations: less generous state incentives, no state incentives, and
eliminating all municipal permit fees. We find that dropping incentives to Step 6 would have
reduced the number of installations by 13 percent in 2014, reducing incentives in half would
have led to a 24 percent drop in installations in 2014, while fully eliminating incentives would
have reduced that number by 47 percent. Based on our estimates, eliminating municipal
permit fees would reduce installations in 2014 by 2.6 percent, although this is likely an
underestimate, for labor costs may also be expected to decrease with expedited permitting
that would likely come along with eliminating municipal permit fees.
Based on our results, we calculate the cost-effectiveness of the CT solar rebate program
to be $135/tCO2 , assuming that natural gas generation is displaced. This is significantly
higher than the roughly $40/tCO2 estimate of the social cost of carbon commonly used by
the U.S. government. Including the benefits from local air pollutants brings the program
closer to being welfare-improving, but consistent with previous studies on California, the
policy is likely to still be difficult to justify based on environmental external costs alone.
Other market failures, such as innovation market failures (van Benthem et al. 2008), must
be significant for the policy to be social welfare-improving.
For firm decision-making, our finding of a price elasticity of -1.76 suggests that consumers
are relatively price sensitive. Firms can use this knowledge for short-run forecasting of the
expected growth of the market. Moreover, our estimated pass-through rate of 84 percent
provides evidence of some level of imperfect competition in this market, consistent with
other studies in the U.S. In an imperfectly competitive market, the price elasticity provides
important guidance to firms for optimal price-setting to maximize profits.
Our results also provide insight into how other policies influence the solar PV market. For
example, we find a highly statistically significant effect of the Solarize grassroots campaigns
on installations. Moreover, we find that when we exclude all Solarize installations from our
28
analysis, our elasticity estimate remains the same, but the variation identifying it is almost
exclusively from the logit specification. This indicates that without Solarize, the data can be
treated as binary rather than count data, a finding due to the fact that the Solarize program
accounts for nearly all installations in a block group-year after the first contract. This may
not be surprising, due to evidence of neighbor or peer effects playing an important role in
the diffusion of solar PV systems (Bollinger and Gillingham 2012; Graziano and Gillingham
2014).
Stepping back, this study of demand in the CT solar PV market highlights the intense
policy effort to promote the technology, which mirrors efforts in many other U.S. states,
European countries, and other countries around the world. As our approach is applicable
to evaluating many solar PV incentive programs, we view further policy analyses in similar
settings as a promising area for future research.
Acknowledgments. The authors would like to thank the Connecticut Green Bank for providing the data used in this analysis. We also thank Bryan Bollinger, CG Dong, Robert Mendelsohn, Corey Lang, Steve Sexton, and numerous seminar audiences for helpful discussions.
Finally, we acknowledge funding from US Department of Energy award DE-EE0006128.
29
2
0
3
500
4
5
price (2014$/watt)
installations
1500
1000
6
7
2000
Figure 1: Trends in Installations and Average Post-Incentive System Price
2008
2009
2010
2011
year
Solarize installations
price
2012
2013
2014
non-Solarize installations
30
0
2000
Frequency
4000
6000
8000
Figure 2: Histogram of the Count of Installations in the Full Sample
0
5
10
15
Number of installations
31
20
0
500
Frequency
1000
1500
Figure 3: Histogram of the Count of Installations in the Truncated Subsample
0
5
10
Number of installations
Solarize installations
15
Non-Solarize installations
32
20
Table 1: Residential Solar Investment Program Timeline
Incentive Type
EPBB/HOPBI
Start Date
Incentive Design:
≤ 5 kW
> 5 kW and ≤ 10 kW
> 10 kW and ≤ 20 kW
PBI
Start Date
Incentive Design:
≤ 10 kW
> 10 kW and ≤ 20 kW
Step 1
Step 2
Step 3
Step 4
Step 5
Step 6
March 2, 2012
May 18, 2012
Jan 4, 2013
Jan 6, 2014
Sept 1, 2014
Jan 1, 2015
$2.450/W
$1.250/W
-
$2.275/W
$1.075/W
-
$1.750/W
$0.550/W
-
$1.250/W
$0.750/W
-
$0.800/W
$0.800/W
$0.400/W
$0.675/W
$0.675/W
$0.400/W
March 2, 2012
May 18, 2012
April 1, 2013
Jan 6, 2014
Sept 1, 2014
Jan 1, 2015
$0.300/kWh
-
$0.300/kWh
-
$0.225/kWh
-
$0.125/kWh
-
$0.180/kWh
$0.600/kWh
$0.080/kWh
$0.060/kWh
Source: The Connecticut Green Bank
33
Table 2: Summary Statistics for the Full Sample
Variable
Number of PV installations
System capacity (kW)
Post-incentive system price ($/W)
Solarize campaign
Population density (per km2 )
Median household income (in 1,000$/year)
Median age
% population above 25 with some college or college degree
% population above 25 with graduate or professional degree
% Republican voters
% Democrat voters
Incentive level ($/W)
Electrical/wiring contractor wage ($/week)
Mean
St. Dev.
Min
Max
0.4279
6.8354
4.2364
0.1277
828.1274
89.376
42.6891
46.8744
18.45
22.0437
33.9395
2.7632
1196.368
1.0884
1.9613
1.956
0.3338
1232.887
42.2584
7.0734
9.9543
12.2112
7.8358
10.4337
1.6322
148.4933
0
0.7
0.0171
0
0
2.499
14.8
0
0
3.69
16.92
0.6824
778.3313
20
22.1
20.5824
1
21784
250.001
80.6
81.203
80.7786
51.14
72.3
6.4879
1651.527
Notes: All variables have 10,150 observations. All dollars in 2014 dollars.
Table 3: Summary Statistics for the Subsample with Positive Installations
Variable
Number of PV installations
System capacity (kW)
Post-incentive system price ($/W)
Solarize campaign
Population density (per km2 )
Median household income (in 1,000$/year)
Median age
% population above 25 with some college or college degree
% population above 25 with graduate or professional degree
% Republican voters
% Democrat voters
Incentive level ($/W)
Electrical/wiring contractor wage ($/week)
Mean
St. Dev.
Min
Max
1.6476
7.2204
3.8164
0.3103
621.817
93.2097
43.4978
47.6733
19.6522
22.8521
32.5627
2.4171
1171.435
1.5976
2.595
1.5613
0.4627
952.9294
41.6419
6.7226
9.4368
12.0778
7.3166
9.3532
1.7173
153.3369
1
0.7
0.0171
0
0
11.838
17.6
0
0
3.69
16.92
0.6824
778.3313
20
22.1
19.0565
1
14530.95
250.001
79.2
77.8978
80.7786
50.19
71.78
6.4879
1651.527
Notes: All variables have 2,636 observations. All dollars in 2014 dollars.
34
Table 4: Monte Carlo Simulation Results
Parameters
Specification
Logit CF
Tr. Poisson GMM
Poisson hurdle
Linear GMM
Poisson CF
True
Value
δ1 = -0.1
δ2 = -0.1
n/a
n/a
n/a
Elasticity
Estimated
Mean
Bias
-0.0967
-0.1002
n/a
-0.2806
-0.1177
0.0033
-0.0002
n/a
n/a
n/a
MSE
True
Value
0.0216
0.0051
n/a
n/a
n/a
η1 = -0.0868
η2 = -0.1975
η = -0.2843
η = -0.2843
η = -0.2843
Implied
Value
Bias
-0.0839
-0.1970
-0.2810
-0.3086
-0.2959
0.0029
0.0005
0.0033
-0.0243
-0.0116
Notes: See Appendix C for details about construction of the simulated data. Output is based on
5,000 replications. Sample means and parameter values, averaged over the 5,000 replications,
are used to compute elasticity under each specification. Elasticity values in the hurdle model
are calculated using the formulas in Section 5.4. The elasticity in the linear model is derived
ˆ
from ηˆ = δy¯w¯ . The elasticity in the Poisson model is derived from ηˆ = δˆw.
¯
35
Table 5: Primary Estimation Results
Linear
Poisson
Variable
Hurdle
Logit
Trun. Poisson
2SLSi
(1)
CFii
(2)
CFii
(3)
GMMi
(4)
Price
-0.0515**
(0.0227)
-0.475***
(0.0617)
-0.466***
(0.0632)
-0.654***
(0.182)
Solarize
0.870***
(0.188)
0.713***
(0.171)
0.019
(0.202)
1.476***
(0.234)
Pop. density
0.00001
(0.00001)
-0.00019**
(0.00009)
-0.00021*
(0.00012)
0.00085
(0.00106)
Income
0.00112
(0.00079)
-0.00012
(0.00185)
-0.00219
(0.00217)
-0.00081
(0.00454)
Age
-0.00263
(0.00272)
-0.0139**
(0.00609)
-0.0130**
(0.00582)
0.00226
(0.0156)
% (some) college
0.00009
(0.00128)
0.00180
(0.00432)
0.00046
(0.00592)
0.0361***
(0.0139)
% grad/prof degree
-0.00040
(0.00182)
-0.00563
(0.00473)
0.00356
(0.00571)
-0.0127
(0.0132)
% Republican
0.0444**
(0.0226)
0.0266
(0.0400)
0.0119
(0.0436)
0.1400
(0.0887)
% Democrat
0.00887
(0.01141)
-0.0177
(0.0292)
-0.0326
(0.0290)
-0.0178
(0.0919)
BG FE
Time trend
Instruments
yes
yes
yes
yes
yes
yes
yes
yes
yes
yes
yes
yes
-0.512**
(0.2246)
-2.012***
(0.2615)
-1.463***
(0.1982)
-0.295***
(0.0819)
10,150
10,150
10,150
2,636
Price elasticityiii
Observations
Notes: Dependent variable is number of residential PV installations. Unit
of observation is block group-year. BG FE refers to block group fixed
effects. The time trend is a quadratic in year. The instruments used for
all IV specifications are the EPBB/HOPBI state financial incentives given
to the installers and the county electrician/wiring wage rate. p < 0.1 (*),
p < 0.05 (**), p < 0.01 (***).
i
Clustered standard errors at the town level in parentheses.
Block bootstrapped standard errors (200 replications), clustered at the
town level, in parentheses.
iii
Standard errors for the price elasticity obtained by the delta method.
ii
36
Table 6: Price Elasticity under Different Specifications
Robustness Checks
Baseline
Model Specification
I
II
III
IV
-1.463***
(0.1982)
-1.469***
(0.1972)
-0.132
(0.6451)
1.460***
(0.1950)
-1.959***
(0.2658)
Tr. Poisson GMMii
-0.295***
(0.0819)
-0.295***
(0.0819)
-0.105
(0.0891)
-0.286***
(0.0797)
-0.102***
(0.0398)
Combined elasticityiii
-1.758***
(0.2145)
-1.764***
(0.2135)
-0.236
(0.6512)
-1.746***
(0.2106)
-2.061***
(0.2687)
Missing price proxy
BG FE
Time trend
Year FE
Demographics/voting
Solarize included
average price
yes
yes
no
yes
yes
highest price
yes
yes
no
yes
yes
average price
yes
no
yes
yes
yes
average price
yes
yes
no
no
yes
average price
yes
yes
no
yes
no
Logit CF
i
Notes: Dependent variable is number of residential PV installations. Unit of observation is block
group-year. All other variables are the same as in Table 5. Standard errors are obtained by the
delta method. p < 0.1 (*), p < 0.05 (**), p < 0.01 (***).
i
Block bootstrapped standard errors (200 replications), clustered by town.
Robust standard errors, clustered by town.
iii
Standard errors of combined elasticity coefficients obtained assuming independence of the data
generating processes.
ii
37
Table 7: Pass-through of Rebates
Variable
I
Rebate level
0.159***
(0.044)
Solarize
-0.554***
(0.054)
Town FE
Time trend
yes
yes
Observations
5,070
Notes: Dependent variable is pre-rebate price.
Specification is a linear least squares regression.
The sample includes only solar contracts with
EPBB/HOPBI rebates. Each contract is a separate observation. Town FE refers to town fixed
effects. The time trend is a quadratic in week.
Standard errors are clustered by town. p < 0.1
(*), p < 0.05 (**), p < 0.01 (***).
Table 8: Policy Simulations
Baseline
Small Reduction
Moderate Reduction
Phase-out
Permit Fees
Year
2014
2015
2014
2015
2014
2015
2014
2015
2014
2015
New installations
Added capacity (MW)
1974
15.43
2026
15.84
1716
13.42
1758
13.75
1510
11.81
1545
12.08
1047
8.18
1069
8.35
2025
15.83
2080
16.26
-258
-13.08
-2.01
-268
-13.21
-2.09
-464
-23.51
-3.62
-481
-23.72
-3.76
-927
-46.99
-7.25
-957
-47.26
-7.49
51
2.58
0.40
54
2.65
0.42
∆ installations
% change in installations
∆ capacity (MW)
Notes: Simulations use a price elasticity estimate of -1.76 from Table 5 and a pass-through rate of 84 percent, as
estimated in Table 7. The “Small Reduction” counterfactual decreases state incentives down to the Step 6 level
(see Table 1) in January 2014. The average rebate in this counterfactual is $0.675/W. The “Moderate Reduction”
counterfactual reduces the subsidy to half of its average 2014 level of $0.935/W in January 2014. The “Phase-out”
counterfactual entirely removes EPBB incentives in January 2014. The “Permit fees” counterfactual entirely removes
municipal permit fees for solar installations in January 2014, reducing average system costs by $0.05/W. Further
details on the counterfactual simulation methodology are in Appendix G.
38
Appendix A
Consistency of the Fixed Effects Truncated Poisson
Estimator
Y i is a vector of outcome variables, such that Y i ∈ Y ⊂ RT+ and each element Yit is
independently distributed with truncated Poisson distribution. Let X it ∈ X ⊂ RP and δ 2 ∈
T
P
∆ ⊂ RP . Let g0 (y|x, n) denote the true conditional distribution of Y i given X i and
Yit .
t=1
If our model is correctly specified, then, for some δ 02 ∈ ∆ ⊂ RP , g(· |x, n, δ 02 ) = g0 (· |x, n),
for all x ∈ X and n ∈ N . We now show that the conditional maximum likelihood estimator,
obtained from the maximization of (16), is a consistent estimator of δ 02 .
Lemma A.1. Let Q(δ 2 ) = E[log g(Y i |X i , ni , δ 2 )] and QN (δ 2 ) =
1
N
N
P
log g(Y i |X i , ni , δ 2 ).
i=1
Under the following assumptions:
A1. ∆ is a compact set
A2. For each (y, x, n) ∈ Y × X × N , log g(y|x, n, · ) is a continuous function on ∆
A3. For all x ∈ X and n ∈ N , Q(δ 2 ) 6= Q(δ 02 ) if δ 2 6= δ 02
P
A4. For all x ∈ X , n ∈ N , and δ 2 ∈ ∆,
g(y|x, n, δ 2 ) = 1
y∈Y
A5. The Uniform Weak Law of Large Numbers holds
A6. The sequence {ci }N
i=1 is bounded, i.e., limN →∞ cN < ∞,
p
δˆ2,CM LE →
− δ 02 , where δˆ2,CM LE = arg max QN (δ 2 ).
δ 2 ∈∆
Proof. From A1 and A2, it follows that the problem max Q(δ 2 ) always has a solution. A3
δ 2 ∈∆
implies that this solution is unique. Using A4 and Jensen’s inequality, it can easily be shown
that Q(δ 2 ) is maximized at δ 02 . From A1 and A2, we also know that the problem max QN (δ 2 )
δ 2 ∈∆
p
always has a solution. Let δˆ2,CM LE = arg max QN (δ 2 ). Under A5, |QN (δˆ2,CM LE )−Q(δ 02 )| →
−
δ 2 ∈∆
p
0, by the Uniform Weak Law of Large Numbers. Therefore, δˆ2,CM LE →
− δ 02 . Note that A6
is needed to ensure that the sequence of unknown parameters that we do not estimate,
c1 , ..., cN , does not contain a high number of extremely large values, which could result in
inconsistency of δˆ2,CM LE .
However, if X is not exogenous, CMLE is no longer consistent. We next show that, in
such a setting, the GMM estimator derived from (24) is a consistent estimator of δ 02 .
0
Lemma A.2. Suppose Z it ∈ Z ⊂ RQ , where Q > P . Let ψ(Z i , δ 2 ) = Z i ξ i and E[ψ(Z i , δ 02 )] =
N
0
P
0
1
0. Furthermore, let G(δ 2 ) = [E[ψ(Z i , δ 2 )] Ξ [E[ψ(Z i , δ 2 )] and GN (δ 2 ) = N [ψ(Z i , δ 2 )
i=1
39
ˆ
Ξ
1
N
N
P
ˆ is a symmetric positive semidefinite weight matrix and Ξ is a
[ψ(Z i , δ 2 ) , where Ξ
i=1
symmetric and positive definite matrix. Under the following assumptions:
A1. ∆ is a compact set
A2. For each z ∈ Z, ψ(z, · ) is a continuous function on ∆
A3. For all z ∈ Z, ψ(z, δ 2 ) 6= ψ(z, δ 02 ) if δ 2 6= δ 02
p
ˆ→
A4. Ξ
− Ξ
A5. The Uniform Weak Law of Large Numbers holds,
p
δˆ2,GM M →
− δ 02 , where δˆ2,GM M = arg min GN (δ 2 ).
δ 2 ∈∆
Proof. A1 and A2 imply that the problem min GN (δ 2 ) always has a solution. Let δˆ2,GM M
δ 2 ∈∆
= arg min GN (δ 2 ). By definition, we know that δ 02 ∈ ∆ solves the problem min G(δ 2 ). By
δ 2 ∈∆
δ 2 ∈∆
A3 and because Ξ is a positive definite matrix, δ 02 is a unique solution to min G(δ 2 ). Then,
δ 2 ∈∆
p
p
using A4 and A5, |GN (δˆ2,GM M ) − G(δ 02 )| →
− 0. Therefore, δˆ2,GM M →
− δ 02 .
40
Appendix B
Monotonicity of m(λ)
In order to ensure that m−1 (· ) is a one-to-one function, we need to show that m(λ) is
monotonic over the relevant range of λ values. In what follows, we prove that, as long as λ
is positive, the function m(λ) is strictly increasing.
0
Lemma B.1. For any λ > 0, m (λ) > 0.
h
i
h λ
i
0
e −1−λ
e−λ
Proof. Dropping all subscripts for simplicity, m (λ) = λ1 − 1−e
m(λ) = λ(e
−λ
λ −1) m(λ).
Note that, by the properties of the truncated Poisson model, m(λ) = Et (Y ) > 0 for all
λ. Furthermore, λ > 0, which implies that λ(eλ − 1) > 0. Hence, we need to ensure that
eλ − 1 − λ is either positive or negative over the relevant range of λ.
0
Let h1 (λ) = eλ and h2 (λ) = λ + 1 and note that h1 (0) = h2 (0). Also note that h1 (0) =
0
0
0
h2 (0) = 1, while, for any λ > 0, h1 (λ) > 1 = h2 (λ). Hence, h1 (λ) > h2 (λ) for all λ > 0,
0
implying that m (λ) > 0 for all λ > 0, which is the relevant range of λ in a truncated Poisson
model.
41
Appendix C
Monte Carlo Simulations
We use the following model in our Monte Carlo simulations:
(
ιit =
1
0
if lit > κit ,
if lit ≤ κit ,
exp(αi + δ1 wit + uit )
,
1 + exp(αi + δ1 wit + uit )
κit ∼ U nif [0, 1],
lit =
(C1)
(C2)
(C3)
(C4)
yit ∼ trP o(λit ), if ιit = 1,
(C5)
λit = exp(βi + δ2 wit + vit ),
(C6)
wit = τ1 αi + τ2 βi + πzit + ρ1 uit + ρ2 vit + eit ,
(C7)
vit , (uit , vit , eit ) zit ,
|=
uit
|=
yit = 0, if ιit = 0,
uit ∼ N (0, σu ), vit ∼ N (0, σv ), eit ∼ N (0, σe ),
(C8)
where i = 1, ..., N, and t = 1, ..., T . Note that wit is correlated both with the error terms uit
and vit , as well as with the individual-specific parameters αi and βi , with the magnitude of
correlation determined by parameters ρk and τk , k ∈ {1, 2}, respectively. The variable zit is
exogenous to the model and is used to instrument for wit . The parameter π measures the
strength of this instrument.
The data generation process is as follows. First, we draw each αi and βi from a uniform
distribution with support [0, 1]. Next, we generate a random variable zit ∼ U nif [0, 5]. After
drawing the random error terms uit , vit , and eit from the distributions shown in (C8), we
plug them into (C7), (C6), and (C2) to obtain wit , λit , and lit , respectively. Finally, we
proceed to generate the outcome variable in each of the two stages of the hurdle model. We
draw κit from the distribution shown in (C3) and then follow the decision rule in (C1) to
generate ιit . We set yit to equal zero for all observations where ιit = 0, as indicated in (C4).
Then, if ιit = 1, yit is drawn from the truncated Poisson distribution in (C5). To obtain the
truncated Poisson draws, we follow Cameron and Trivedi (2013) by drawing from a Poisson
distribution with parameter λit and then replacing each zero draw with another draw from
the same distribution until all draws are positive.
The parameters to be chosen are δ1 , δ2 , τ1 , τ2 , π, ρ1 , ρ2 , σu , σv , σe , N , and T . We set
δ1 = δ2 = -0.1, τ1 = τ2 = 0.2, π = 0.8, ρ1 = ρ2 = 0.5, σu = σv = σe = 0.5, N = 25, and
T = 4. We replicate the estimation 5,000 times, each time re-drawing different uit , vit , and
eit .
42
Appendix D
First-stage Instrumental Variable Results
To demonstrate the strength of our instruments, Table D1 shows the results of a first-stage
regression of the post-incentive PV system price per W on all instruments for both the
full and truncated sample. The coefficients on incentives are highly statistically significant.
A joint F-test of statistical significance of the three excluded instruments provides a test
statistic of 845.20 in the full sample and 86.36 in the truncated sample.
Table D1: First-stage Instrumental Variable Regression Results
Variable
Full Sample
Truncated Sample
-0.966***
(0.0245)
-0.822***
(0.0626)
-0.00064***
(0.00023)
0.00041*
(0.00026)
-0.297***
(0.0404)
-0.398***
(0.0785)
3.5×10−5
(3.6×10−5 )
0.00030*
(0.00018)
Income
-0.00027
(0.00083)
-0.00333
(0.00203)
Age
-0.00127
(0.00269)
-0.00705
(0.00678)
% (some) college
-0.00013
(0.00193)
0.00423
(0.00523)
% grad/professional
0.00101
(0.00211)
0.00871
(0.00693)
% Republican
-0.0268*
(0.0159)
0.0714*
(0.0378)
% Democrat
-0.0401***
(0.0093)
0.0340
(0.0351)
BG FE
Time trend
yes
yes
yes
yes
F-statistic
845.20
86.36
Observations
10,150
2,636
Incentive level
Wage rates
Solarize
Pop. density
Notes: Dependent variable is post-incentive PV system
price per W. Unit of observation is block group-year.
Specification is a linear least squares regression. BG FE
refers to block group fixed effects. The time trend is a
quadratic in year. p < 0.1 (*), p < 0.05 (**), p < 0.01
(***).
43
Appendix E
Non-instrumented Regression Output
Table E1: Non-instrumented Regression Results
Linear
Poisson
Variable
Hurdle
Logit
Truncated Poisson
OLSi
(1)
MLEii
(2)
CMLEii
(3)
CMLEi
(4)
Price
0.0527***
(0.00765)
0.0908***
(0.0195)
0.1830***
(0.0264)
-0.1355***
(0.0388)
Solarize
0.858***
(0.182)
0.913***
(0.207)
0.107
(0.232)
1.554***
(0.216)
3.9×10−6
(8.8×10−6 )
-0.00020**
(8.1×10−5 )
-0.00021*
(0.00011)
0.00037
(0.00083)
Income
0.00120
(0.00076)
8.7×10−5
(0.00180)
-0.00125
(0.00187)
0.00193
(0.00374)
Age
-0.00217
(0.00255)
-0.01140*
(0.00607)
-0.00861
(0.00549)
0.00146
(0.0138)
% (some) college
0.00047
(0.00128)
0.00306
(0.00421)
0.00297
(0.00529)
0.0267***
(0.0101)
% grad/prof degree
-6.8×10−5
(0.00166)
-0.00384
(0.00475)
0.00570
(0.00525)
-0.01259
(0.00971)
% Republican
0.0518**
(0.0224)
0.0306
(0.0388)
0.0352
(0.0439)
0.0910
(0.0685)
% Democrat
0.0168
(0.0110)
0.0102
(0.0307)
0.0102
(0.0307)
-0.0016
(0.0813)
BG FE
Time trend
Instruments
yes
yes
no
yes
yes
no
yes
yes
no
yes
yes
no
0.521**
(0.0757)
0.385**
(0.0828)
0.573***
(0.0829)
-0.061***
(0.0175)
10,150
10,150
10,150
2,636
Pop. density
Price elasticityiii
Observations
Notes: Dependent variable is number of residential PV installations. Unit of
observation is block group-year. BG FE refers to block group fixed effects. The
time trend is a quadratic in year. p < 0.1 (*), p < 0.05 (**), p < 0.01 (***).
i
Clustered standard errors at the town level in parentheses.
Block bootstrapped standard errors (200 replications), clustered at the town level,
in parentheses.
iii
Standard errors of price elasticity coefficients obtained by the delta method.
ii
44
Appendix F
Including Third-party-owned Systems
Table F1: Regression Results with Purchased and Third-party-owned Systems
Linear
Poisson
Variable
Hurdle
Logit
Truncated Poisson
2SLSi
(1)
CFii
(2)
CFii
(3)
GMMi
(4)
Price
0.0566
(0.0513)
-0.1580*
(0.0906)
-0.252**
(0.126)
-0.0402
(0.0919)
Solarize
1.467***
(0.249)
0.518***
(0.107)
-0.263
(0.242)
0.865***
(0.112)
TPO
0.181***
(0.043)
0.129*
(0.067)
0.449***
(0.086)
-0.144*
(0.0825)
Pop. density
0.00004***
(0.00001)
-0.00012***
(0.00005)
-0.00016*
(0.00008)
0.00069*
(0.00042)
Income
0.00291***
(0.00080)
0.00031
(0.00125)
-0.00132
(0.00190)
0.00268
(0.00262)
Age
0.00060
(0.00238)
-0.01296***
(0.00402)
-0.01404***
(0.00540)
0.00167
(0.00938)
% (some) college
0.00045
(0.00134)
-0.00019
(0.00348)
0.00058
(0.00504)
0.00231
(0.00515)
% grad/prof degree
-0.00316**
(0.00153)
-0.00983***
(0.00360)
0.00168
(0.00537)
-0.0221***
(0.00783)
% Republican
0.08337**
(0.03621)
0.00628
(0.03906)
0.01226
(0.04505)
0.0357
(0.0443)
% Democrat
0.02291
(0.01916)
0.00944
(0.02131)
-0.00925
(0.02878)
0.0707**
(0.0308)
BG FE
Time trend
Instruments
yes
yes
yes
yes
yes
yes
yes
yes
yes
yes
yes
yes
0.312
(0.2827)
-0.609*
(0.3490)
-0.647**
(0.3224)
-0.025
(0.0577)
13,447
13,447
13,447
4,513
Price elasticityiii
Observations
Notes: Dependent variable is number of residential PV installations. Unit of observation is block group-year. TPO measures the fraction of new installations that
are third-party-owned. All other variables are the same as in Table 5. p < 0.1 (*),
p < 0.05 (**), p < 0.01 (***).
i
Clustered standard errors at the town level in parentheses.
Block bootstrapped standard errors (200 replications), clustered at the town level,
in parentheses.
iii
Standard errors of price elasticity coefficients obtained by the delta method.
ii
45
Appendix G
Inputs for the Policy Simulations
Table G1 lists the values of the main parameters and variables used in our policy simulations in Section 7. As an estimate of system costs in 2015, we assume that the trend of
declining module and inverter costs continues. In order to derive the predicted change in
pre-incentive average system costs (i.e., the sum of installer reported module, inverter, labor, and permitting costs) during 2015, we extrapolate the trend from 2008 to 2014. More
specifically, we calculate the year-to-year percentage change in the average pre-incentive cost
per W (calculated as the ratio of the average cost to the average system size) over this time
period and fit a simple exponential curve through these percentage changes. We then take
the extrapolated 2015 value as our forecasted cost per W in 2015. This value is shown in
the top row of Table G1.
The average 2014 values of all installation-related variables are obtained directly from
our core dataset. In addition, the average EPBB/HOPBI rate is derived using the mean
system capacity value in 2014 at Step 4 and Step 5 incentive levels and taking the average of
the two resultant rates. Lastly, our data contains information on the structure of municipal
solar PV system permit fees. While some CT towns charge a flat fee per installation, in most
towns the fee depends on the capacity of the installation. In those towns, we use our system
size data for each installation in 2014 to obtain the fee for that installation. We average
across all installations to derive an average town fee per installation and then average across
all towns. After dividing the resultant number by average system capacity, we obtain the
average town fee per W, reported in the last row of Table G1.
Table G1: Input Values for Policy Simulations
Parameter/Variable
Value
Change in average system cost from Jan 1, 2015 to Dec 31, 2015 (%)
Number of block groups
Average number of installations per block group in 2014
Average system capacity (kW) in 2014
Average system price ($/W) in 2014
Average rebate rate ($/W) in 2014
Average town permit fee ($/W) in 2014
-5.30
1450
1.361
7.819
2.946
0.935
0.0518
46
References
Andersen, E. B. (1970), “Asymptotic Properties of Conditional Maximum-Likelihood
Estimators”, Journal of the Royal Statistical Society, Series B, 32, 283–301.
Barbose, G., N. Darghouth, S. Weaver, and R. Wiser (2013),“Tracking the Sun
VI: An Historical Summary of the Installed Price of Photovoltaics in the United States
from 1998 to 2012”: Berkeley, CA: Lawrence Berkeley National Laboratory.
Barbose, G., S. Weaver, and N. Darghouth (2014),“Tracking the Sun VII: An Historical Summary of the Installed Price of Photovoltaics in the United States from 1998 to
2013”: Berkeley, CA: Lawrence Berkeley National Laboratory.
Berry, S. T., J. A. Levinsohn, and A. Pakes (1995), “Automobile Prices in Market
Equilibrium”, Econometrica, 63, 841–889.
Blundell, R., R. Griffith, and F. Windmeijer (2002), “Individual Effects and Dynamics in Count Data Models”, Journal of Econometrics, 108, 113–131.
Bollinger, B. and K. Gillingham (2012), “Peer Effects in the Diffusion of Solar Photovoltaic Panels”, Marketing Science, 31, 900–912.
(2014), “Learning-by-doing in Solar Photovoltaic Installations”, Yale University
Working Paper.
Borenstein, S. (2012), “The Private and Public Economics of Renewable Electricity Generation”, Journal of Economic Perspectives, 26, 67–92.
Cameron, A. C. and P. K. Trivedi (2013),“Regression Analysis of Count Data”: Cambridge: Cambridge University Press, 2nd edition.
CEFIA (2013), “Sunrise New England: Final Project Report”, Clean Energy Finance and
Investment Authority Report.
Chamberlain, G. (1980), “Analysis of Covariance with Qualitative Data”, Review of Economic Studies, 47, 225–238.
Dong, C., R. Wiser, and V. Rai (2014), “Incentive Pass-through for Residential Solar
Systems in California”, Lawrence Berkeley National Laboratory Report.
Gallop, R. J., R. H. Rieger, S. McClintock, and D. C. Atkins (2013), “A Model
for Extreme Stacking of Data at Endpoints of a Distribution: Illustration with W-Shaped
Data”, Statistical Methodology, 10, 29–45.
Gillingham, K., B. Bollinger, and H. Staver (2014a), “The Role of Information and
Social Interactions on Solar Adoption Decisions and Equilibrium Prices: Evidence from a
Field Experiment”, Yale University Working Paper.
47
Gillingham, K., H. Deng, R. Wiser, N. Darghouth, G. Barbose, G. Nemet,
V. Rai, and C. Dong (2014b), “Deconstructing Solar Photovoltaic Pricing: The Role
of Market Structure, Technology, and Policy”, Yale University Working Paper.
Gowrisankaran, G., S. Reynolds, and M. Samano (2015), “Intermittency and the
Value of Renewable Energy”, University of Arizona Working Paper.
Graff Zivin, J., M. J. Kotchen, and E. T. Mansur (2014), “Spatial and Temporal
Heterogeneity of Marginal Emissions: Implications for Electric Cars and Other ElectricityShifting Policies”, Journal of Economic Behavior and Organization, 107A, 248–268.
Graziano, M. and K. Gillingham (2014), “Spatial Patterns of Solar Photovoltaic System Adoption: The Influence of Neighbors and the Built Environment”, Journal of Economic Geography, forthcoming.
Gupta, S. and S. Park (2009), “A Simulated Maximum Likelihood Estimator for the
Random Coefficient Logit Model Using Aggregate Data”, Journal of Marketing Research,
46, 531–542.
Henwood, K. (2014), “Subsidy Pass-through in Residential Solar Markets”, University of
Chicago Working Paper.
Hughes, J. E. and M. Podolefsky (2015), “Getting Green with Solar Subsidies: Evidence from the California Solar Initiative”, University of Colorado at Boulder Working
Paper.
IAWG (2013),“Inter-Agency Working Group on Social Cost of Carbon, Technical Support
Document: Technical Update of the Social Cost of Carbon for Regulatory Impact Analysis
under Executive Order 12866”: Washington, DC.
Imbens, G. and J. M. Wooldridge (2007),“What’s New in Econometrics? Lecture 6:
Control Function and Related Methods”, NBER Summer Institute.
Jiang, R., P. Manchanda, and P. E. Rossi (2009), “Bayesian Analysis of Random
Coefficient Logit Models Using Aggregate Data”, Journal of Econometrics, 149, 136–148.
King, G. (1988), “Statistical Models for Political Science Event Counts: Bias in Conventional Procedures and Evidence for the Exponential Poisson Regression Model”, American
Journal of Political Science, 32, 838–863.
Kitazawa, Y. (2012), “Hyperbolic Transformation and Average Elasticity in the Framework
of the Fixed Effects Logit Model”, Theoretical Economics Letters, 2, 192–199.
Majo, M. C. and A. H. O. van Soest (2011),“The Fixed-Effects Zero-Inflated Poisson Model with an Application to Health Care Utilization”, Discussion Paper 2011-083,
Tilburg University, Center for Economic Research.
Min, Y. and A. Agresti (2005), “Random Effect Models for Repeated Measures of ZeroInflated Count Data”, Statistical Modelling, 5, 1–19.
48
Mullahy, J. (1986), “Specification and Testing of Some Modified Count Data Models”,
Journal of Econometrics, 33, 341–365.
(1997), “Instrumental-Variable Estimation of Count Data Models: Application to
Models of Cigarette Smoking Behavior”, Review of Economics and Statistics, 79, 586–593.
Muller, N. Z., R. Mendelsohn, and W. Nordhaus (2011), “Environmental Accounting for Pollution in the United States Economy”, American Economic Review, 101,
1649–1675.
Neelon, B., P. Ghosh, and P. F. Loebs (2013), “A Spatial Poisson Hurdle Model for
Exploring Geographic Variation in Emergency Department Visits”, Journal of the Royal
Statistical Society: Series A, 176, 389–413.
Neyman, J. and E. L. Scott (1948), “Consistent Estimates Based on Partially Consistent
Observations”, Econometrica, 16, 1–36.
Olley, G. S. and A. Pakes (1996), “The Dynamics of Productivity in the Telecommunications Equipment Industry”, Econometrica, 64, 1263–1297.
Petrin, A. and K. Train (2010), “A Control Function Approach to Endogeneity in
Consumer Choice Models”, Journal of Marketing Research, 47, 3–13.
Pohlmeier, W. and V. Ulrich (1995), “An Econometric Model of the Two-Part Decisionmaking Process in the Demand for Health Care”, Journal of Human Resources, 30,
339–361.
Rogers, E. and S. Sexton (2014), “Effectiveness of Subsidies for Residential Rooftop
Solar Adoption: Lessons from California”, North Carolina State University Working Paper.
Sallee, J. (2011), “The Surprising Incidence of Tax Credits for the Toyota Prius”, American Economic Journal: Economic Policy, 3, 189–219.
Santos Silva, J. M. C. and F. Windmeijer (2001), “Two-part multiple spell models
for health care demand”, Journal of Econometrics, 104, 67–89.
Shaw, S., R. Fahey, and P. Solomon (2014), “Progress Report: CEFIA Residential
Solar Investment Program”, The Cadmus Group, Inc. Report.
Terza, J. V. (1998), “Estimating Count Data Models with Endogenous Switching: Sample
Selection and Endogenous Treatment Effects”, Journal of Econometrics, 84, 129–154.
van Benthem, A., K. Gillingham, and J. Sweeney (2008), “Learning-by-Doing and
the Optimal Solar Policy in California”, The Energy Journal, 29, 131–151.
Windmeijer, F. (2008),“GMM for Panel Data Count Models”, in L. M´aty´as and P. Sevestre
eds. “The Econometrics of Panel Data”, Berlin Heidelberg: Springer-Verlag.
49
Windmeijer, F. and J. M. C. Santos Silva (1997), “Endogeneity in Count Data Models: An Application to Demand for Health Care”, Journal of Applied Econometrics, 12,
281–294.
Winkelmann, R. (2004), “Health Care Reform and the Number of Doctor Visits - An
Econometric Analysis”, Journal of Applied Econometrics, 19, 455–472.
Wooldridge, J. (2002),“Econometric Analysis of Cross Section and Panel Data”: MIT
Press.
Yang, S., Y. Chen, and G. M. Allenby (2003), “Bayesian Analysis of Simultaneous
Demand and Supply”, Quantitative Marketing and Economics, 1, 251–275.
50